NBER WORKING PAPER SERIES

RESOURCE WINDFALLS, POLITICAL REGIMES, AND POLITICAL STABILITY

Francesco CaselliAndrea Tesei

Working Paper 17601http://www.nber.org/papers/w17601

NATIONAL BUREAU OF ECONOMIC RESEARCH1050 Massachusetts Avenue

Cambridge, MA 02138November 2011

We are very grateful to Antonio Ciccone for many discussions and to Tim Besley, Silvana Tenreyro,and seminar participants at LSE and UPF for comments. Caselli gratefully acknowledges the supportof CEP, and Banco de España, the latter through the Banco de España Professorship, and the hospitalityof CREI, where the project was initiated. The views expressed herein are those of the authors and donot necessarily reflect the views of the National Bureau of Economic Research.

NBER working papers are circulated for discussion and comment purposes. They have not been peer-reviewed or been subject to the review by the NBER Board of Directors that accompanies officialNBER publications.

© 2011 by Francesco Caselli and Andrea Tesei. All rights reserved. Short sections of text, not to exceedtwo paragraphs, may be quoted without explicit permission provided that full credit, including © notice,is given to the source.

Resource Windfalls, Political Regimes, and Political StabilityFrancesco Caselli and Andrea TeseiNBER Working Paper No. 17601November 2011JEL No. E0,F0

ABSTRACT

We study theoretically and empirically whether natural resource windfalls affect political regimes.We document the following regularities. Natural resource windfalls have no effect on the politicalsystem when they occur in democracies. However, windfalls have significant political consequencesin autocracies. In particular, when an autocratic country receives a positive shock to its flow of resourcerents it responds by becoming even more autocratic. Furthermore, there is heterogeneity in the responseof autocracies. In deeply entrenched autocracies the effect of windfalls on politics is virtually nil, whilein moderately entrenched autocracies windfalls significantly exacerbate the autocratic nature of thepolitical system. To frame the empirical work we present a simple model in which political incumbentschoose the degree of political contestability by deciding how much to spend on vote-buying, bullying,or outright repression. Potential challengers decide whether or not to try to unseat the incumbent andreplace him. The model uncovers a reason for the asymmetric impact of resource windfalls on democraciesand autocracies, as well as the differential impact within autocratic regimes.

Francesco CaselliDepartment of EconomicsLondon School of EconomicsHoughton StreetLondon WC2A 2AEUNITED KINGDOMand NBERf.caselli@lse.ac.uk

Andrea TeseiUniversitat Pompeu FabraCarrer Ramon Trias Fargas 25-27BarcelonaSpainandrea.tesei@upf.edu

1 Introduction

Looking at the historical experiences of specific countries it seems uncontroversial that an abundanceof natural resources can shape political outcomes. Few observers of Venezuela, Nigeria, Saudi Arabia,and many other resource-rich countries would take seriously the proposition that political developments inthese countries can be understood without reference – indeed without attributing a central role – to thesecountries’ natural wealth. Yet, the mechanisms through which natural-resource abundance affects politicsfrustrate attempts to identify simple generalizations, with resource-rich countries displaying great variationsin measures of autocracy and democracy, and political stability. For example, Saudi Arabia and Nigeriaboth feature a strong tendency towards autocracy but the former is extraordinarily stable while the latterhas experienced nine successful coups since independence (and many unsuccessful ones). Venezuela seems togo back and forth between democracy and autocracy, with swings that closely follow the price of oil, whileof course Norway appears to be safely and stably democratic irrespective of the oil price.

In this paper we use a large panel of countries to document the following regularities. Natural resourcewindfalls have no effect on the political system when they occur in democracies. However, windfalls havesignificant political consequences in autocracies. In particular, when an autocratic country receives a positiveshock to its flow of resource rents it responds by becoming even more autocratic. Importantly, there is het-erogeneity in the response of autocracies. In deeply entrenched autocracies the effect of windfalls on politicsis virtually nil. It is only in moderately entrenched autocracies that windfalls exacerbate the autocraticnature of the political system. Hence, our evidence generalizes casual observation: windfalls have little orno impact in democracies (the Norways) or very stable autocracies (Saudi Arabia), but change the politicalequilibrium in more unstable autocracies (Nigeria, Venezuela).

To reach these conclusions we measure natural-resource windfalls as changes in the price of a country’sprincipal export commodity. We argue that such changes are exogenous to a country’s political system.1

While total resource exports may depend on political developments, the identity of a country’s main exportcommodity (e.g. oil v. gold) is unlikely to depend on politics. Similarly, the vast majority of countriesindividually account for a relatively small share of world output in their principal export commodity, soit is unlikely that political changes there will have an important effect on prices. Our main measure ofpolitical institutions is from the Polity IV database. Crucially for our analysis this is a continuous measurethat varies from extreme autocracy to perfect democracy, so it allows us to condition the analysis on infra-marginal differences in the degree of autocracy/democracy, as well as to capture the effects of windfalls oninfra-marginal changes in autocracy/democracy. As this variable captures the extent to which the politicalsystem is open to competition, we sometimes refer to our measure of autocracy/democracy as a measure of“political contestability.”

In order to motivate our empirical analysis, and facilitate the interpretation of the results, we open thepaper with a simple model of endogenous determination of political contestability. In our model there is agoverning elite that has complete control of the flow of income from natural resources, and decides whetherand how much of it to invest in what we call “self-preservation activities.” These range from the mild (e.g.direct or indirect vote-buying) to the extreme (violent repression of the opposition). At the same time, apolitical entrepreneur outside the ruling elite decides whether or not to challenge those in power and try to

1In the empirical analysis we address the issue of large producers with the potential to influence world prices, and find thatour results are not affected by these economies.

1

replace them. This simple game generates endogenously two possible political “modes” : free and fair politicalcompetition (recognizable as democracy), where the elite essentially allows challenges to occur on a relativelylevel-playing field, and the political entrepreneur chooses to compete for power; and a “repression” modewhere the elite invests some of the resources deriving from natural resources in self-preservation activities,without however succeeding in completely deterring challenges.

In our basic model, the key determinant of the regime that is selected as an equilibrium is the amount ofrevenue accruing to the government from natural resources. This enters the ruling elite’s decision problemin two ways: it is part of the payoff from staying in office, as political survival implies that the current eliteremains in control of future revenues; and it also enters the budget constraint, as it is the principal source offunding for self-preservation activities, such as vote-buying or political repression. At low levels of resourceincome, the incentive to engage in self-preservation spending is relatively low, as the future “pie” to holdon to is small. Democracy is the outcome. At higher levels the future benefits from holding on to powerare sufficiently large that the government shifts to autocracy. Crucially, the larger the pie, the more theincumbent finds it optimal to spend on self-preservation, so the degree of autocracy is increasing in the sizeof the resource rents.

One prediction of the model is that political contestability is non-linearly related to resource abundance.Resource-poor countries will be democratic, while resource-rich ones will be autocratic, and the level ofautocracy will be increasing in the amount of resource rents. We show that this simple cross-country relationis consistent with the data, when resource abundance is measured in terms of commodity exports. Howeverfor reasons we discuss later this is not a compelling test. We therefore note that another prediction of themodel is that resource-poor countries (democracies) will not experience changes in political contestabilityfollowing (small) resource shocks, while resource-abundant countries will. Furthermore, in the model, therate of decline in political contestability following changes in resource rents is decreasing in the initial levelof resource rents (and hence in the initial level of autocracy). This is due to an assumption of decreasingreturns in self-preservation spending by the incumbent government. Hence, the model also predicts that inautocracies the effect of windfalls is decreasing in the extent to which the autocracy is entrenched. Thispredicted heterogeneity in response between democracies and autocracies, as well as within autocracies, isthe focus of our empirical work.

The threshold levels of resource income that cause the shift from one political regime to the other dependon parameters that may vary across countries. In a very simple extension to our basic model, in particular,the thresholds depend on a parameter that could be interpreted as the return offered by the markets on thehuman capital of defeated politicians. If former politicians can look forward to decent returns on their talentin the market, the range of values of natural wealth for which the ruling elite accepts free and fair challengesis (potentially much) wider than in places where politics is the only road to riches. In this way, the modelcan potentially also explain cases, such as Norway, where great natural-resource wealth is associated withdemocracy.

The paper continues as follows. In the next subsection we briefly review the relevant literature. Section2 presents the model and Section 3 presents data and results. Section 4 concludes.

2

1.1 Related Literature

An important literature in political science studies the relationship between resource abundance anddemocratic/autocratic institutions using predominantly comparative case studies or cross-country varia-tion [e.g. Ross (2001a, 2001b, 2009), Ulfelder (2007), Collier and Hoeffler (2009), Alexeev and Conrad(2009) and Tsui (2010)]. While there is some heterogeneity in the conclusions this literature tends to reach,the evidence in these studies points to a negative relationship between resource abundance and democ-racy/democratization, consistent both with our model and the circumstantial cross-sectional evidence wealso present below. However, we argue that identification of causal effects can be achieved with greaterconfidence using within-country variation, and this is the basis for the core of our empirical evidence.

A recent literature narrowly focused on windfalls from oil uses within-country evidence. Haber andMenaldo (2010) and Wacziarg (2009) find no evidence that oil windfalls lead to greater autocracy. Oneconcern with the Haber and Menaldo (2010) study is that its measure of oil revenue, partly based onoil production, is potentially endogenous to democratic change, while a possible concern with Wacziarg’sanalysis is that it uses the world oil price for all countries, meaning that there is no possibility to control forcorrelated time effects across countries. Bruckner, Ciccone, and Tesei (2011) find a positive coefficient onoil-price shocks interacted with the share of net oil exports in GDP in a regression for movements towardsdemocracy. They do not condition on whether the country is initially a democracy or an autocracy, nor dothey examine heterogeneous responses within autocracies.2

On the theoretical front, Acemoglu, Robinson, and Verdier (2004) present a model of autocratic rulewhere, as in our model, natural-resource rents affects both the value of holding power and the resourcesavailable to the incumbent to protect himself. In many ways our model is a much simplified version oftheirs. However, they focus on a dichotomous outcome (democracy v. autocracy) so their analysis has nopredictions on how the effect of windfalls will vary within autocracies. Also, in their framework democraciesare absorbing states by construction, whereas we derive this endogenously. Finally, somewhat more subtly, intheir framework the main results are derived as a consequence of resource windfalls relaxing the kleptocratbudget constraint, whereas in our model the main mechanism is the variation in the value of staying inpower. Haggard and Kaufman (1997) and Geddes (1999) also stress the role of the budget constraint ofpolitical incumbents.

Still fairly closely related are studies of how economic changes other than resource windfalls affect democ-racy/autocracy, and how resource windfalls affect political outcomes other than democracy/autocracy.

Prominent in the first vein is the long tradition, stretching back at least to Lipset (1959) of studies linkingchanges in incomes to changes in political institutions. Acemoglu and Robinson (2001, 2006) develop modelsin which temporarily low income lowers the opportunity cost of challengers to the status quo, leading todemocratic transitions in autocracies and reversals in democracy. Our model differs in that the effect ofincome changes depends not only on whether the political system is initially autocratic or democratic, butalso on infra-marginal heterogeneity in the degree of initial autocracy. Furthermore, to make the modelspeak more directly to the effects of commodity booms we model the economic mechanism not as a changein the opportunity cost of challengers but as a change in the reward of holding political power. Furthermore,

2A possible interpretation of the result in Bruckner, Ciccone, and Tesei (2011) is that, since the oil share is highly correlatedwith autocracy, their oil-share/oil-price interaction operates as a rough proxi for our autocracy/oil-price interaction. The resultsare therefore consistent, as in both cases they imply a lesser movement towards autocracy in more entrenched autocracies.

3

we model changes in political regime not as a dichotomous transitions towards democracy but as continuouschanges in the degree to which the regime represses political contestability. Because of these differences,our predictions also differ markedly, as we predict no response in democracies and infra-marginal changes inautocracies.

Many authors have investigated empirically the causal relationship between income and democracy [e.g.Barro (1999), Epstein et al. (2006), Ulfelder and Lustik (2007), Glaeser, Ponzetto, and Shleifer (2007),Acemoglu, Johnson, Robinson, and Yared (2008), Bruckner and Ciccone (2009), and Burke and Leigh(2010).] As discussed, we focus not on generic income changes but more specifically on windfalls associatedwith commodity price shocks. Because natural-resource booms typically translate into direct windfalls intothe hands of political elites these shocks may have very different political consequences than other sources ofincome shocks. In fact, the literature on the natural resource curse casts doubt on the premise that resourcewindfalls are aggregate-income increasing [e.g. Sachs and Warner (2001)]. Burke and Leigh (2010) do usecommodity price changes as instruments for income changes, so their work is more closely related. Theyfind insignificant effects of commodity-driven income changes on political regimes. Their focus, however, ison dichotomous variables measuring the onset of large changes towards autocracy or democracy. Instead,in keeping with the spirit of our model, we study changes in autocracy/democracy as a continuous variable.Furthermore, Burke and Leigh do not condition the effect of commodity price changes on whether the countrywas initially democratic or autocratic, much less on infra-marginal differences in the initial level of politicalcontestability. Finally, as already mentioned, in Burke and Leigh the effect of windfalls is mediated by theireffect on income changes, while we estimate the direct effect of the windfall. For the reasons mentionedabove there may be reasons to prefer a reduced-form specification.

As for the literature studying the effects of resource windfalls on political outcomes other than democ-racy/autocracy, a very incomplete list includes the theoretical studies of Tornell and Lane (1998), Balandand Francois (2000), and Torvik (2002), all of whom study theoretically the consequences of windfalls for rentseeking, and Leite and Weidmann (1999), Tavares (2003), Sala-i-Martin and Subramanian (2003), Dalgaardand Olsson (2008) and Caselli and Michaels (2011) that present corresponding empirical evidence (whererent seeking is usually measured though proxies of corruption). Caselli and Coleman (2006) examine theoret-ically the consequences of resource abundance for ethnic conflict, and Besley and Persson (2010) for politicalviolence.3 Cabrales and Hauk (2009) study the impact of resource windfalls on incumbents’ probability ofre-election, though their main focus is on the consequences for human-capital accumulation.

2 Natural Resources and Political Outcomes

2.1 Model

The setting is a discrete-time infinite-horizon economy which generates, in every period, a constant flowof consumption goods A from the exploitation of natural resources. Interpretations of A include: the flowof royalties and other fees paid to the government by international extracting companies for the right tooperate in the country; profits of state-owned corporations engaged in drilling and mining; rents generated

3The mechanism in Caselli and Coleman (2006) and Besley and Persson (2010) is similar to ours: fiscal windfalls increase thevalue of holding power leading to grater incentives to engage in exploitation of others, repression and (for the non-incumbentgroup), fight back. However, initial political institutions are taken as exogenous, while they are endogenous in our model.

4

by the international distribution of domestic cash crops by state-controlled marketing boards; or other rentslinked to cash-crops exports due to discrepancies between official and market exchange rates. We will referto A as “resource rents.”

The economy is populated by a very large number of infinite-lived agents. In every period one agent,which we term the “incumbent,” has complete control of these flows, in the sense that he can decide howto allocate them between different uses. One should think of the incumbent as the individual or group ofindividuals who has de facto control of the government - and is hence in receipt of the resource rents. In ademocracy this would be the President and his collaborators (in presidential systems), or the leadership ofthe governing parties (in parliamentary systems). In autocracies this would be the autocrat, his family, andhis close associates. Aside from the benefits associated with control of the resource rent A, an incumbentalso receives a flow of “ego rents” Θ. Assuming that there are additional benefits (both psychological andmaterial) from holding political power is realistic and indeed standard in the literature.

In every period another agent (not the incumbent) is randomly selected by nature to be the “potentialchallenger.” The potential challenger is given an opportunity to try to replace the incumbent. In particular, ifthe potential challenger decides to attempt to unseat the incumbent, the attempt will succeed with probabilityp. p is endogenous as we discuss shortly. In a democracy the potential challenger could be interpreted asthe person with the best chance to win an electoral context against the incumbent president/party. In anautocracy it could be the agent best placed to successfully lead a coup or a popular uprising against theruling clique. The assumption that in every period there is only one potential challenger is not importantfor the results but simplifies the analysis. For simplicity of presentation and again without loss of generalitywe also assume that potential challengers are drawn without replacement (i.e. each agent gets at most onechance to challenge) and that deposed incumbents never get a chance to challenge subsequent incumbents.The potential challenger has an outside option represented by the present value of his activities outsidepolitics, which we denote Π.

As mentioned the incumbent must allocate the resource rent among possible uses. One use of theresource rents is what we call “self-preservation.” Self-preservation spending is any spending that reduces theprobability that a challenge succeeds (conditional on a challenge occurring). Hence, if Bt is self-preservationspending, the probability of a successful challenge is p(Bt), with p′(Bt) < 0. Our interpretation of self-preservation spending is as a catch-all for all activities the government engages in in order to subvert theoutcome of the political-selection process in his favor. It includes vote-buying and patronage spending, buyingand/or bullying and intimidation of the media, and outright repression and persecution of opponents. Thehigher is B, the more aggressive and draconian the tactics employed. Hence, we think of variation in B

as capturing infra-marginal variation in the efforts exerted by those currently in power to subvert the rulesof the game in their favor, with greater values of B being associated with greater autocracy. By the sametoken, we think of B = 0 as the situation where the incumbent accepts to be challenged on a “free and fair”basis. In sum, we interpret countries with B = 0 as “democracy” and countries with B > 0 as displayingvarying levels of autocracy. Since B also affects a potential challenger’s chances of taking over we will alsorefer to B as a measure of political contestability.

In order to obtain crisp results, we need to pick a functional form for p(B). We use

p(B) = Ωe−δB ,

5

where Ω ∈ (0, 1) and δ > 0 are exogenous parameters. Hence, self-preservation spending is subject todecreasing returns, with p(0) = Ω - implying that a challenger can never be absolutely certain of success -and p(B) > 0 for all B - implying that an autocrat can never be absolutely sure of successfully withstandinga challenge. These features are important but seem sensible.

The portion of A not spent on self-preservation is spent on another activity, which we call “consumption”and denote by Ct. Ct provides a direct utility flow to the incumbent, so that his total utility flow in periodt is Ct + Θ. Obviously one interpretation of Ct is resources appropriated by the incumbent and his cliquefor personal enrichment - the infamous “Swiss bank accounts.” But in general Ct could be interpreted as anaggregate of all the spending that provides satisfaction to the incumbent and hence, possibly, it could includepublic spending on schools, hospitals, etc., if the incumbent is partially altruistic or derives satisfaction fromdoing a “good job.”

The restrictive assumption is that the components of Ct do not affect p or Π. If the public is lesstolerant of corrupt politicians, then we might expect the components of Ct that represent self-enrichmentto enter p positively. If the public rewards competent politicians, we should expect the components of Ctthat represent public spending to enter p negatively, in the tradition of Barro (1973). In addition, publicspending in infrastructure, human capital, and other growth-promoting public goods could improve theoutside option of potential challengers by improving opportunities in the private economy (or increasing thecost of recruiting supporters). Hence, these components of Ct could increase Π. We abstract from theseissues in order to get simple results, but see Caselli and Cunningham (2009) for a detailed discussion.4 Weare also implicitly assuming that there is no scope for government borrowing, though as we discuss belowthis assumption could easily be replaced by an assumption that incumbents face an upward-sloping supplycurve for borrowing, without qualitative changes in the results.

The series of events within each period is the following. First the incumbent allocates the period rentsA between self-preservation Bt and consumption Ct. Next nature picks a potential challenger, and thepotential challenger decides whether or not to try to unseat the incumbent. If yes, then the challenge successwith probability p(Bt). If the challenge succeeds, the challenger becomes the new incumbent. If it fails,the incumbent continues as incumbent, as he does if the challenger foregoes the opportunity to try. Time isdiscounted by all agents at rate β.

2.2 Analysis

We formally analyze the model in the Appendix. Here we offer a heuristic discussion and explain the keyresults.

We focus on Markov Perfect Equilibria (MPE), of which we show there is only one. Given that the onlystate variable is the resource rent A, and this is constant over time, it is immediate that players will follow

4A straightforward extension in the direction of allowing productive public spending would be as follows. Rents are allocatedbetween repression, B, private consumption, C, and productive public spending, G, and the probability of successful challengeis decreasing in both B and G: p(B) = Ωe−δB−γG. It is immediate to show that in this case the incumbent never uses bothtools at the same time. In particular, if δ > γ the incumbent only uses repression, while if δ < γ he only uses productive publicspending. Hence, one interpretation of the model is that we focus on the case δ > γ. Another interpretation is that the relativemagnitudes of γ and δ vary across countries, perhaps for cultural, geographic, or geostrategic reasons. In countries where δ > γthe analysis in the rest of this section applies. Countries with γ > δ will obviously be democracies, as we defined democracyas a country with B = 0. Furthermore, in such countries shocks to natural-resource rents will have no impact on B. Hence,we recover the same empirical prediction as in the baseline model, namely that we should observe no systematic response ofpolitical institutions to resource shocks in democracies.

6

stationary strategies, namely the incumbent will set the same value of B in every period, while the potentialchallenger will either always challenge or never challenge.

We begin by establishing the conditions for equilibria where the challenger always challenges. In such anequilibrium, the value of being an incumbent at the beginning of any period is

V (A,B′) =Θ +A−B′

1− β [1− p(B′)],

where B′ is the equilibrium level of self-preservation spending. In every period the incumbent receivesego rents Θ and consumes resource rents net of self-preservation spending A − B. This flow utility isappropriately discounted by taking into account time preferences β and the fact that in each period theprobability of “political death” is p(B′). Note that for simplicity we have normalized the continuation valueafter losing office to 0.

One condition for an equilibrium with challenges is that the level of self-preservation spending must befeasibly optimal from the point of view of the current incumbent. The current incumbent’s problem is

maxBΘ +A−B + β [1− p(B)]V (A,B′)

s.t..B ≥ 0

B ≤ A

In choosing B the incumbent trades off the short-term decline in consumption with the improved probabilityof surviving until next period and enjoying the continuation value of office. The feasibility constraints say thatself-preservation spending cannot be negative and cannot exceed the resources available to the incumbent.

Now define b(A,B′) as the solution to the above problem. In an equilibrium, b(A,B′) must be a fixedpoint, or

b(A,B′) = B′.

In the appendix we show that this fixed-point problem has a unique solution. In particular, there existsa value of A, A0, such that the solution is at the corner B′ = 0 for A ≤ A0, while for A > A0 B

′ is theinterior solution to the problem above. We call this interior solution B∗(A). B∗(A) is increasing, concave,and satisfies B∗(A0) = 0. The intuition for this result is simple, and can be illustrated with reference tothe incumbent’s problem above. The marginal cost of extra preservation spending is constant and equal to1. The marginal return is −p′(B)βV (A,B′), i.e. the improvement in survival probabilities times the valueof surviving. Since the value of surviving is increasing in A, there can be sufficiently low values of A suchthat the incumbent renounces all self-preservation efforts. On the other hand, if A is sufficiently large, theincumbent spends (increasing) amounts on self-preservation. The equilibrium amount of self-preservation isthe one that equalizes marginal cost and marginal benefit.5

The threshold value A0 is given by

A0 =1− β(1− Ω)− βΩδΘ

βΩδ,

5We show in the appendix that the other constraint, B ≤ A, is never binding.

7

and is therefore decreasing in the “ego rents” from holding office. Intuitively, the larger the ego rents, theless the level of resource rents required to make the incumbent feel that incumbency is valuable enough toinvest resources in protecting it. The technology of political replacement also affects A0. In particular, ahigher productivity of self-preservation spending, δ, makes the incumbent more willing to exert efforts inthis direction, lowering the threshold for autocratic behavior.

As mentioned above we think of B = 0 as akin to the idea of “free and fair” political competition, andhence as democracy. Since democracy is the observed equilibrium outcome in many countries, we assumethat there exists a region of the parameter space where it occurs. Formally,

Parametric Assumption (PA) 1:A0 > 0.

A second condition for an equilibrium where the challenger challenges is that challenging is optimal giventhe level of self-preservation efforts exerted by the incumbent. If the equilibrium incumbent strategy is B,the challenger decides to challenge if

p(B)βV (A,B) > Π. (1)

The left hand side is the expected utility of challenging. This is equal to the time-discounted value ofbeginning next period as the incumbent, times the probability that the challenge will succeed. Note thatwe are implicitly assuming that the challenger experiences no flow utility in the period of the challenge(conditional on a challenge occurring). This could easily be relaxed without any change in results. Alsonote that for simplicity we normalize the value of a defeated challenger to 0. We discuss the implicationsof relaxing this assumption below. The right hand side is the (certain) utility from not challenging, i.e. theoutside option Π.6

Since the value of holding office is increasing in A, condition (1) is satisfied for A if it is satisfied forA = 0. In turn, the condition is satisfied for A = 0 if the following parametric assumption holds.7

Parametric Assumption (PA) 2

Π <βΩΘ

1− β(1− Ω).

Note that for A = 0 the incumbent chooses democracy. If PA2 did not hold incumbents would face nochallenges in democracies. This would be counterfactual so PA2 seems like a plausible assumption. Thesimple interpretation of PA2 is that the ego rents from office are sufficiently attractive relative to private lifeto make potential challengers willing to try their luck at politics (when there are no resource rents and thecountry is a democracy).

A final requirement for an equilibrium where the challenger challenges is that the incumbent does not tryto completely deter a challenge in the current period. The deviation that does so is the one that satisfies (1)with equality. Call Bc(A) such a deviation. We show that there exists a level of A, A, such that Bc(A) > A

for all A < A. This says that “resource poor” incumbents cannot afford the level of preservation spendingthat would be required to completely deter challenges. Only when A is sufficiently large can an incumbent

6Note that Π depends on β. In particular, if π is the flow utility in the private sector then Π = π/(1− β).7To see that PA1 and PA2 are mutually consistent notice that PA1 can be rewritten as

βΩΘ

1− β(1− Ω)<

1

δ

8

achieve complete control of his destiny. The value of A is given by

A =1δ

logβΩΘ

Π(1− β).

This is increasing in the ego rents. Larger ego rents mean that potential challengers are less easily deterred,i.e. the required investments in self-preservation are larger, and therefore unaffordable for a broader rangeof values of A. Similarly, A is decreasing in the opportunity cost of challenging and in the productivity ofspending.

For values of A ≥ A deviating to a strategy of complete deterrence is feasible, and the question iswhether the deviation is preferred. It turns out that this depends on whether log(δΠ) + 1 ≥ 0 – in whichcase the deviation is preferred – or log(δΠ) + 1 < 0, in which case the incumbent sticks to the “interior”(non-deterring) amount of preservation spending. The intuition is that both δ and Π reduce the cost of fulldeterrence, the former by increasing the productivity of preservation spending, and the latter by making thechallenger more easily convinced thanks to a better outside option.

For reasons to be discussed shortly, we assume that even when a deviation is feasible the incumbent willnot deviate from the “interior” strategy. Formally,

Parametric Assumption (PA) 3:

log(δΠ) + 1 < 0.

This leads to the following summary of the discussion so far.Lemma 1. Under PA2 and PA3 a MPE where the challenger challenges exists for all A. If A ≤ A0 then

B = 0 (democracy). If A > A0 then B = B∗(A) (autocracy).We can now turn to the conditions for a MPE where the challenger is deterred. In this equilibrium the

incumbent invests an amount B(A) that solves

p(B)βV (A, B) = Π,

where V (A, B) is now the value of incumbency when the challenger does not challenge. B(A) is increasingand concave. By definition of B(A) the challenger is deterred. Not surprisingly it turns out that the policyis feasible if A ≥ A, but it is preferred by the current incumbent to a one-period deviation to the optimal“interior” level of B if PA 3 holds. Hence, we have the following result.

Lemma 2.Under PA3 there is no MPE where the challenger is deterred.The reason for imposing PA3 are largely dictated by events in the Middle East and North-Africa of

early 2011. Even regimes that were the byword for stability and entrenchment have appeared unable todiscourage attempts to unseat them, irrespective of the large amounts of resources (from oil or from foreignaid) available to their rulers. PA3 rules out the possibility of complete deterrence.8

Note that Lemmas 1 and 2 imply that the MPE is unique. This gives rise to the following conclusion.Conclusion. In the unique MPE equilibrium, resource poor countries are democracies, while resource

rich countries are autocracies. In autocracies, spending on self-preservation is an increasing and concavefunction of the resource rents.

8If we were to replace PA2 with its opposite , and assumed A0 ≤ A then we would have three types of political regimes:democracies (B = 0 for A ≤ A0); unstable autocracies (B = B∗ for A0 < A ≤ A); and stable autocracies (B = B for A < A).While this set of outcomes may have seemed anecdotally appealing until recently we think this is no longer the case.

9

This result says that for values of the resource rent that are sufficiently small the value of staying in officeis limited, and does not justify spending on self-preservation. Hence, resource poor countries will tend to bedemocratic. For higher values of resource rents the incumbent finds it optimal to exert efforts to remain inpower, and does so up the point where the extra improvement in the expected value of staying in office isequal to the marginal cost of resources spent on self-preservation. Figure 1 depicts the equilibrium amountof self-preservation spending as a function of A.

To get us closer to our empirics we now consider the following thought experiment. Suppose that atsome date the value of A unexpectedly increases by a (small) amount dA, and all agents expect it to remainconstant at this value for the indefinite future (this is all consistent with rational expectations if A is believedto be a random walk). Then we obviously have

dB = 0 for A ≤ A0

dB = B∗′ for A0 < A

Hence, in resource-poor countries marginal increases in resource rents lead to no political change. However, incountries with non-negligible resource rents, further windfalls induce an increase in self-preservation spending.In particular for intermediate values of the rent flow the incumbent becomes keener to stay in office, andhence increases his efforts in this direction. For even larger initial levels of the resource flow, the incumbentfinds that the required amount of spending needed to deter challengers goes up, and must correspondinglyincrease it. Because B∗ is a concave functions of A, the response of self-preservation spending is decreasingin the resource flow over this range.

Combining the two sets of results on the level of B and the change of B as functions of the initial level of A,it is also possible to recast the latter set of results as conditioned on the initial level of democracy/autocracy.In particular, as we have noted, for low levels of A countries tend to be democratic. This implies that indemocracies, marginal changes in the flow of resource rents have no effect on the political equilibrium. Forlarger values of the resource rent, countries are autocracies. Hence, we find that in autocracies, marginalchanges in the flow of resource rents make the political equilibrium more autocratic. Furthermore, the degreeof tightening of the autocratic screws is variable. Clearly the concavity of B∗ with respect to the initiallevel of A also carries through to the relationship between the change and the initial level of B. Hence,in autocracies, the increase in autocracy following an increase in resource revenues is diminishing in theinitial level of autocracy. For reasons we discuss below, the core empirical work in the paper is based on thepredictions of this paragraph.

To fully appreciate the potential for the model to map into real-world outcomes it is essential to notethat the threshold A0 depends on parameters that are potentially country-varying. For example, a declinein the effectiveness of self-preservation spending δ or in the ego rents Θ shift the autocracy threshold A0

to the right. In other words, countries with greater cultural, geographical, historical, or external resistanceto autocracy – all features that should map into a lower value of δ – or countries where the same factorsdictate that the balance between the privileges and the responsibilities of political power weighs the lattermore (low Θ), will remain in democratic mode for a wider range of values of A. This way, the model canperhaps be seen as consistent with cases of high A associated with free and fair democracy, such as Norway.

Possibly one limitation of the model above is that we impose a balanced-budget constraint. It shouldbe fairly obvious that the mechanism highlighted in the previous section will continue to work even if the

10

government can tap into foreign financial markets to finance self-preservation spending. All that is needed isthat the government faces an upward sloping supply curve of funds. relative to the model presented above,the marginal cost of self-preservation spending would be increasing, rather than a constant. There would stillbe a threshold analogous to A0 for autocracy, and for higher values of A there would still be a unique interioroptimal amount of self-preservation spending, increasing in A. A sufficient condition for self-preservationspending beyond A0 to still be concave would be that the supply curve for foreign loans is convex, which isvery likely to be the empirically relevant case.

3 Evidence

3.1 General strategy

The main result of the paper is a highly non-linear relationship between resource income A and self-preservation efforts B, as depicted in Figure 1. In principle, there are three possible approaches to try toidentify this relationship empirically. We discuss the three approaches and explain why only one, which wediscuss last, is likely to generate compelling evidence. In discussing the three approaches we assume we havegood measures of A and B. In the next section we discuss the data in detail.

Given a measure of B the first plan that comes to mind (Plan A) is to try to get a measure of A andthen use non-linear methods to directly estimate the function in Figure 1 using cross-country data in levels.There are at least two problems with this approach. First, is the well-rehearsed vulnerability of cross-countryrelationships to omitted variable bias. There may be plenty of hard-to-account-for factors correlated bothwith the volume of resource rents and the political system. Second, as discussed at the end of the previoussection the autocracy threshold A0 is likely to be country specific. Appropriate identification would thereforerequire explicitly modelling the dependence of A0 on hard-to-measure country specific factors. The resultswould likely be fairly untransparent and inconclusive.

Plan B investigates the relationship between A and B within countries, or, equivalently, in differences,conditioning on the initial level of A. Looking at the effects of changes in A on changes in B eliminatestime-invariant confounding country-specific factors that bias inference in levels. Country fixed effects canbe added to control for country-specific trends in democracy/autocracy and time effects can be added tocontrol for global trends. Hence, plans B largely sidesteps the first of the identification issues affecting PlanA. However, because it conditions on the initial level of A, Plan B still requires an estimate of country-specificautocracy thresholds A0, so it is still unsatisfactory.

Plan C, like plan B, estimates the relationship in differences, but instead of conditioning on the initiallevel of A it conditions on the initial level of B. Our theoretical results say that countries to the left of theautocracy threshold are democracies so we can infer that if a country is a democracy it is to the left of itsA0. We therefore expect no effect of changes in A on changes in B in democracies. We also know fromthe model that countries to the right of A0 are autocracies, and the further to the left they are the moreautocratic they are. Hence, we can infer that autocracies are to the right of A0, and the more autocraticthey are the further to the right they are. We therefore expect that the effect of changes in A on changes inB is positive in autocracies, the less so the more autocratic the initial position. This plan largely sidestepsboth the problem of omitted factors in levels and the country-specificity of the autocracy threshold.

11

3.2 Data

We construct a measure for B from the variable Polity2 in the Polity IV database [Marshall and Jaggers(2005)]. Polity2 is widely used in the empirical political-science literature as a measure of the position ofa country on a continuum autocracy-democracy spectrum [e.g. Acemoglu, Johnson, Robinson, and Yared(2008), Persson and Tabellini (2006, 2009); Besley and Kudamatsu (2008); Bruckner and Ciccone (2009)]. Itaggregates information on several building blocks, including political participation (existence of institutionsthrough which citizens can express preferences over policies and leaders), constraints on the executive, andguarantees of civil liberties both in daily life and in political participation, as evaluated by Polity IV coders.Polity2 varies continuously from -10 (extreme autocracy) to +10 (perfect democracy). Note, therefore,that polity2 is an inverse measure of B.9 We follow the convention in the vast majority of the literaturethat interprets negative values of polity2 as pertaining to autocracies and positive ones to democracies [e.g.Persson and Tabellini (2006, 2009); Besley and Kudamatsu (2008); Bruckner and Ciccone (2009), Olken andJones (2009), Epstein et al. (2006)]. Nevertheless we discuss alternative thresholds in Section 3.4.

To map the Polity Score into a proxy for B we make the following assumption:

Polityit = α− f(Bit) + εit, (2)

where Bit is our variable of interest, f is a monotonic function with f(0) = 0, α > 0 is a constant, and εit

is an i.i.d. error with zero mean. These assumptions imply that when the government does not attempt tosubvert in its favour the political process (B = 0) the polity measure tends to be positive and its variationto depend on factors we do not model. Instead, when the government takes an autocratic stance, the polityvariable is decreasing in the aggressiveness of this stance.

As long as f(B) is not (too) convex, Assumption (2) implies that the polity score will inherit the sameproperties of B in the model. In particular, for values of the polity score associated with democracies(polity> 0, or B = 0) changes in A have no systematic effect on changes in polity score. In autocracies(negative polity, or positive B) increases in A have negative but decreasing effects on changes in the polityscore.

To measure natural-resource windfalls at the country level we proceed as follows. First, for each countryand for each year that data is available we rank all commodities (in the universe of agricultural and mineralcommodities) by value of exports. We then identify each country’s principal commodity as the commoditythat is ranked first in the largest number of years. The export data by commodity, country, and year arefrom the United Nation’s Comtrade data set, which reports dollar values of exports according to the SITC1system, for the period 1962 to 2009. Finally, we match each country’s principal commodity with an annualtime series of that commodity’s world price. All commodity prices are extracted from the IMF IFS dataset,with the exception of Gemstones, Pig Iron and Bauxite, whose price series are obtained from the UnitedStates Geological Survey.

9The Polity2 variable is a modification of the basic Polity variable, added in order to facilitate the use of the Polity intime-series analysis. It converts “standardized authority scores”(-88, -77, -66) into conventional polity scores. In particular,“foreign interruptions” (-66) are coded as missing variables, cases of “interregnum and anarchy” (-77) are coded as 0, and“transitions” (-88) are prorated through the time span of the transition. We adjust Polity2 by assigning missing values to casesof interregnum and anarchy, to avoid the misleading representation of autocracies progressing toward democracy in periodsof anarchy. In section 3.4 we investigate the robustness of our results to alternative adjustments of “standardized authorityscores.”

12

We identify a change in A in country i as a change in the price of country i’s principal commodity.As both the identity of a country’s principal commodity and its price in international markets are largelyexogenous to the country’s political outcomes we think this measure allows for clean identification of thecausal effects of resource windfalls (we investigate robustness to dropping the largest producers below).

We study changes over the period 1962-2009. Our baseline sample consists of 131 countries with in-formation on both principal-commodity export shares and polity2 scores. There are 32 distinct principalcommodities in this sample. The most frequent are oil, which is the principal commodity in 30 countriesand coffee (11 countries). Table 1 gives the list of these principal commodities and their distribution amongcountries.

Summary statistics are presented in Table 2. ∆Polity is the one-year difference in Polity2, whileAvg∆Price is the average growth rate in the price of the principal commodity over a three-year window(we discuss timing issues below). AvgShare is the average over time of the value of exports of a country’sprincipal commodity as a share in GDP. Countshare Princ indicates the number of years the principal com-modity has been the principal export, while Countshare is the total number of years in which commodityshares are available. Some of the notable features in the data are the huge variation in the polity2 score(spanning the entire set of possibilities in all years) and the secular trend towards greater democracy. Thetable also shows that principal commodities are ranked first in almost all years in which resource shares areavailable. Finally, the table shows that there is much variation in the measure of resource windfalls.

3.3 Results

Our main empirical results are presented in Table 3. The dependent variable is the one-year change inPolity2. Recall that an increase in this variable means that the country becomes less autocratic (more demo-cratic). In column 1 the explanatory variable is the lagged change in the price of the principal commodity,averaged over the previous three years. Hence, if the change in Polity2 is measured between years t− 1 andt, the change in commodity prices is the average over the years t−4, t−3, t−2, and t−1. We look at laggedchanges in prices to defuse lingering concerns about reverse causation, as well as to allow for possible lagsin the reaction of political actors to economic events. We take averages of price changes over three periodsto reduce the role of extremely transitory shocks as well as measurement error in the explanatory variable.By construction, however, the rolling windows introduce serial correlation in the estimates. To account forthis, we cluster the standard errors at the country level in all regressions, allowing for heteroskedasticityand arbitrary correlation in the error term. We further report on robustness to timing assumptions below.Country and time fixed effects are included here, and in all subsequent specifications.

Column 1 reports estimates for the average effect of resource windfalls, which is negative but not sta-tistically significant. Recall that in our theory the average effect is a weighted average of nil effects indemocracies and negative effects in autocracies, and thus depends on the relative frequency of autocraticand democratic observations. In our sample the number of democracies substantially exceeds the numberof autocracies (2570 versus 2305 observations). It is therefore not surprising that the overall effect is notstatistically significant.

In the remaining columns we test our more detailed predictions. Column 2 looks at the effect of pricechanges in democracies and autocracies separately. This is accomplished by separating out the price-changevariables into two variables: the first is an interaction between the price change and a dummy for autocracy

13

(following the literature convention that identifies autocracies as countries with a negative polity2 score);the second is an interaction between price change and a dummy for democracy (non-negative polity2 ).10 Tobe consistent with the starting date for the price shock implied by our lagging choices we measure the initiallevel of autocracy democracy with a four year lag, or in year t− 4. As predicted by the model, price changesin the principal commodity have a negative impact on the polity2 score in autocracies, i.e. make autocraciesmore autocratic. Instead, they have no significant impact on the polity2 score in democracies.11

Our model not only has predictions for the average effect of resource windfalls in democracies andautocracies, but also on the relative magnitude of the effect depending on the initial value of (resourcerents and hence) the measure of political contestability. In particular, the prediction is that in democraciescommodity-price changes will have no impact not only on average but also for any initial level of polity2.Instead, in autocracies the magnitude of the effect should be increasing in contestability: small in veryaggressive autocracies, and larger as the autocracy takes milder forms. We test this prediction in Column 3,where we add four-year lags of polity2 both, by themselves and interacted with the (autocracy/democracyspecific) price change, the latter being the variable of interest. The conditioning variable has been enteredwith a lag to allow once again for potentially slow responses by political actors. As predicted, in democraciescommodity price changes have no impact at any level of initial polity2, while in autocracies the increase inautocracy following a resource windfall is larger the higher the initial value of polity2, i.e. the less autocraticthe form of government was initially.

The results in Columns 1-3 are based on OLS estimation. In Column 4 we show that the results arevirtually unchanged using System-GMM estimation. System-GMM provides consistent estimates in dynamicpanel data model with fixed effects, by instrumenting the differenced variables that are not strictly exogenouswith all their available lags in levels and differences. This removes the bias introduced in OLS estimatesby the within group transformation, which by construction produces a correlation between the transformedlagged dependent variable and the transformed error term [Nickell (1981), Bond (2002)]. The system GMMresults in Column 4 are very close to the original OLS.

Our main results are illustrated in Figure 2, which plots the estimated effect of a change in the price ofthe principal commodity on the change in polity2, conditional on the initial level of polity2, together with90% confidence bands. In the top panel, we have the average (unconditional) effect, which is negative butinsignificant. In the middle panel we have average effects in democracies and autocracies separately. Theeffect is negative in autocracies and nil in democracies. In the bottom panel we plot the response conditionalon infra-marginal differences in contestability. The increase in autocracies is more severe the milder theinitial level of autocracy.

The estimated coefficients imply that the impact of resource windfalls for a weak autocracy (say, atpolity2 level -2) is more than twice as large as the one for a more consolidated autocracy at polity2 level -6.In the weak autocracy, the long-run effect of resource windfalls implies that a 10% increase in the price of theprincipal commodity reduces the Polity2 score by 1.65 points, or 8% of the domain of Polity2 (which goesfrom -10 to +10). For the more consolidated autocracy, instead, the effect of a 10% increase in the price ofthe principal commodity only reduces the Polity2 score by 0.8 points. An alternative way to put this is that

10It can be easily checked that this is equivalent to including the price change by itself and then an interaction betweenthe price change and, say, a democracy dummy. Our specification makes the interpretation of the coefficients even morestraightforward.

11In Column 2 and elsewhere in the table the coefficients on the level of initial contestability is negative in democracies, andinsignificant in autocracies. This suggests that there is some convergence among democracies but not among autocracies.

14

a weak autocracy like Ecuador (average Polity2 score in autocracy -2) needs a 24% price shock to move toa more consolidated form of autocracy, like Nigeria’s (average Polity2 score in autocracy -6). For Nigeria toexperience a similar 4 points reduction in the Polity2 scale, and become like Saudi Arabia (average Polity2score in autocracy -10), the price increase in the principal commodity should be of 50%.

3.4 Robustness checks

In this section we report a number of robustness checks on our results from the previous subsection. Inparticular, we discuss robustness to: alternative criteria for inclusion in the sample based on (i) importanceof the principal commodity in the economy and (ii) accuracy of the identification of the principal commodity;(iii) focusing on observations away from the lower and upper bounds of polity2 ; (iv) dropping large commodityproducers with the potential of influencing the world price; (v) measuring resource-rent shocks based on abasket of commodities rather than only the principal commodity; (vi) breaking down commodities by type(mineral v. non-mineral; point-source v. diffuse); (vii) alternative ways to treat problematic values of polity2 ;(viii) alternative measures of the outcome variable; (ix) alternative timing structures for the relationshipbetween outcomes and shocks; and (x) alternative thresholds for democracy.

Table 4 checks the robustness of our results to the exclusion of countries whose principal commodityaccounts for only a small share of GDP. For these countries it is unlikely that a price change representsa large windfall, so focusing on a smaller sample with significant principal-commodity share is arguablya better test for our model.12 Columns 1 to 3 exclude countries in the first decile of the average sharedistribution (14 countries, typically modern democracies with a diversified economy); columns 4 to 6 excludecountries in the first quartile (38 countries); and columns 7 to 9 exclude all countries below the medianaverage share (68 countries). Results from baseline sample are confirmed and generally reinforced as weprogressively increase the threshold to be included in the sample. In particular the point estimates for theaverage effect in autocracy (columns 2, 4 and 6) become more negative as we focus on more commoditydependent countries. Also the lagged level of polity2 interacted with the (autocracy specific) price changeremains negative and significant throughout all subsamples, confirming the heterogeneous impact of resourcewindfalls within autocracies.

In Table 5 we show results are robust to the possible sample selection induced by data availability. Assome countries only report few years of export data, their principal commodity may be poorly identified.We address this concern in a number of ways. First, we rank countries by number of years on which it waspossible to identify the principal commodity. We then drop the 25% with fewest years (which turn out to becountries with at most sixteen years of export-share data). The results are in Columns 1 to 3. Columns 4to 6 restrict the sample to countries for which we observe export data for the principal commodity at leastonce before 1986, which is the mid-point of the sample period. In columns 7 to 9 we also include thosecountries that do not have share data before 1986, but whose principal commodity has always been rankedfirst afterwards. In this case it is plausible to assume that the commodity had been important before 1986,even though there is lack of data to confirm it. Our results are robust to all these checks.

Table 6 investigates the robustness of our result on the heterogeneous impact of resource windfalls within12However this benefit should be weighed against the fact that the size of the commodity sector is endogenous. Hence this

exercise reintroduces through the back door of sample selection the endogeneity issues we sought to avoid by focusing on pricechanges. This is why the exercise is a robustness check. Our prefereed approach remains the one in the previous section.

15

autocracies. One potential concern is that such heterogeneity might be driven by the boundedness of thepolity scale. The argument is that observations at the -10 boundary are more constrained in their movementsthan non-boundary observations. In particular, as they can’t go lower than -10, price increases would notresult in institutional changes. We address this concern in a number of ways. In column 1 we restrict thesample to non-negative polity2 changes, so that countries at the -10 boundary are unconstrained in theirmovements. We still find a negative and highly significant heterogeneous effect among autocracies. In column2 we perform a similar exercise, but replace the polity2 change by a dummy variable that takes the valueof 1 if we observe a positive change and 0 otherwise. This weighs all institutional changes equally. Theheterogeneous impact of price variation is also maintained under this specification. Column 3 restricts thesample to all countries that never touched the [-10, +10] boundaries. This is the sample of countries thathad effective free movements in both positive and negative directions. Also in this case, we find a negativeand significant effect among autocracies. Finally in column 4 we exclude all country-year observations atthe [-10, 10] boundaries. Limiting the sample to the unbounded cases provides consistent estimates forcensored regressors [Rigobon and Stoker (2007, 2009)]. The results again confirm the heterogeneity amongautocracies.

In a further effort to probe the role of observations at the -10 boundary, we estimate the heterogenouseffect of price changes non parametrically. We divide all observations into six bins, depending on the valueof polity2, and re-estimate the relationship between changes in polity2 and changes in principal-commodityprices separately for each of these bins (always including country and year fixed effects). The six bins arefor polity2 values [-10,-8], [-7,-6], [-5,0], [1,5], [6,7], and [8,10], respectively. These bin sizes were chosen tohave as uniform as possible a sample size across bins, while at the same time preserving symmetry between“autocratic” and “democratic” bins. The estimated coefficients and the relative confidence bands (at the90% level) are plotted against the average value of polity2 in each bin in Figure 3. The figure shows thateven in the second bin from the bottom, the effect of price changes is considerably weaker than in the thirdbin. This is important because for observations in this bin the upper bound at -10 does not appear ever tobe binding. To check this, we have calculated, for each initial value of the polity2 variable on the right handside of our main regression, the fraction of (strictly) negative policy changes equal to the distance from thelower bound, on the left hand side. For example, for observations at polity2=-7, we computed the fractionof negative changes equal to -3. The results, reported in Appendix Table 1 (together with the analogousnumbers for positive changes), show that the lower bound at -10 is never binding for changes in any of thefive bins other than the bottom bin.13

In Table 7 we address the plausible concern that current commodity prices are affected by expectationsof future political developments in the main world producers. We therefore exclude from the sample allcountries belonging to OPEC (columns 1 and 2) and those accounting for more than 3% of total worldproduction of their principal commodity (columns 3 and 4).14 Despite the significant drop in sample size, inour key specifications the results on the heterogeneous impact among autocracies remain robust at least at

13It may seem strange that there are strictly negative changes at initial polity2 equal to -10, but remember that the initialvalue is observed with a four year lag. It is thus possible for a country that was at -10 4 years before to have since moved to alevel strictly less than -10 and then have a regression between year t-1 and t.

14We treat Indonesia and Gabon as OPEC countries, as they belonged to the organization for more than half of the sampleperiod. Instead, we exclude Angola and Ecuador, who joined the OPEC only in 2007. Alternative treatments of these countriesdo not alter the results. A list of the major producers by prinicpal commodity , as well as data sources for commodity production,is given in Appendix Table 2.

16

the 10% significance level.Our source of identification for resource windfalls stems from variations in the international price of the

principal commodity. Other authors in this field [Deaton (1999), Ciccone and Bruckner (2010), Besley andPersson (2010)] use instead a country-specific composite price index, weighting commodity prices by the eachcommodity’s share in the country’s total exports. We have not followed this strategy because of concernswith the possible endogeneity of commodity shares (as well as measurement error). However in Table 8we check the robustness of our results to this alternative specification, constructing a country-specific indexbased on commodities in our sample. In column one we weigh price changes by each country’s time averageof the share of that commodity in exports. Because time coverage of the share data varies dramatically overtime, these averages are also computed over very different time periods from country to country. In the othercolumns we follow the far superior practice of using shares in a given year. The downside of this is thatsample sizes shrink significantly as in each year there is a sizable subset of countries for which shares are notobserved. In these experiments, the qualitative patterns of our baseline results are robust, but statisticalsignificance is not always achieved.

In Table 9 we deal with the issue of commodity typology. An important distinction that has been madein the literature is between point source and diffuse natural commodities [Sokoloff and Engerman (2000),Isham et al. (2005)]. The former are believed to foster weaker institutional capacity and induce greaterresistance to democratic reforms than the latter, as they are generally more valuable and easier to controlfor the ruling elite. We therefore expect our theory to apply more strictly to point source countries. Wetake as point source all mineral commodities plus coffee, cocoa, sugar and bananas [agricultural commoditiesidentified as point source in Sokoloff and Engerman (2000), Isham et al. (2005)]. Our data show thatpoint source producers are indeed more autocratic (average polity2 level -0.81) than countries with diffuseprincipal commodities (average polity2 level 3.13). A mean comparison test rejects the null hypothesis ofmeans equality at the 99% confidence level (t-stat 17.4). Column 2 in Table 9 confirms our baseline results forthe sample of point source producers: the impact of resource windfalls is negative and heterogeneous withinautocracies, while it has no effect in democracies. Column 4 shows instead an average significant effect fordiffuse commodity producers, but no significant heterogeneity. In columns 5 to 8 we consider an alternativeclassification, taking as point source commodities minerals only. Column 6 confirms the results for mineralautocratic countries. Column 8 considers non mineral countries only and displays a negative average relationbetween price and institutional change, with no evidence of heterogeneity in the effect. Altogether, Table 9provides support for our theory in point source producers under both alternative classifications, while it isless conclusive for diffuse commodities producers.

As reported in section 3.2, the polity2 variable codes foreign interruptions as missing variables, cases ofinterregnum and anarchy with a “neutral”score of 0, while transitions are prorated through the time spanof the transition. There exists a general agreement in recent literature on the miscoding of interregnumand anarchy, as the 0 score often produces the wrong representation of autocracies progressing towarddemocracy in periods of anarchy [Bruckner and Ciccone (2010), Burke and Leigh (2010), Plumper andNeumayer (2010)]. The adopted solution consists in assigning missing values to interregnum and anarchyperiods. We have applied the same methodology in this paper. In Table 10 we set all observations pertainingto transition periods to missing and we still get (in Column 3) a significantly heterogeneous effect amongautocracies.

17

One set of robustness checks that did not prove consistent with our baseline results was the use of alterna-tive proxies for our variable B, which in our model represents the self-preservation activities of incumbents.We tried two alternative measures of political repression: the Political Terror Scale (Wood and Gibney, 2010),and the CIRI index of human rights (Cingranelli and Richards, 2008).15 PTS uses data from Amnesty Inter-national and the US State Department. It gives a classification 1-5 from lowest to highest human insecurityand provides a single score into which multiple dimension of abuse have been collapsed. The CIRI indexexplicitly codes four different types of abuses: disappearances; political torture; imprisonment of politicalopponents; killing of political opponents. It then constructs a nine point scale of “physical integrity”basedon the sum of these components. Neither of these measures turned out to be significantly related to resourcewindfalls in our sample. One relevant concern with such measures of repression however is that they onlycapture outcomes. As has been noted by other authors as well, the PTS (but the same can be said of theCIRI Index) “measures actual violations of physical integrity rights more than it measures general politicalrepression. In fact there will be instances in which one government is so repressive that, as a consequence,there are relatively few acts of political violence” (Wood and Gibney, 2010, p. 370). This is to say, mostrepressive countries can score low values of human rights violations as the high expected punishment detersany actions that could trigger overtly repressive acts. This represents a main difference with respect to thepolity2 variable, which attempts to capture not only outcomes but also procedural rules. In addition, polity2aims to include a broader set of dimensions along which political activity can be distorted, beyond physicalrepression. These observations are corroborated by the low correlation between the polity2 scores and thePTS and CIRI scores (0.36 and 0.37, respectively).

Throughout our empirical analysis the main explanatory variable is the lagged change in the price of theprincipal commodity, averaged over the previous three years. This means that institutional changes between1979-1980 are explained by average price changes in 1977-1979; institutional changes between 1980-1981are explained by average price changes in 1978-1980, and so on. The rolling window specification has theclear advantage of smoothing out extreme observations and reducing measurement error, and the resultingserial correlation in the estimates can be dealt with by clustering the standard errors at the country level,allowing for heteroskedasticity and arbitrary correlation in the error term, as we have done. To further checkthe robustness of our results to the timing structure, Table 11 presents estimates using three years non-overlapping windows. This reduces the sample size by two thirds, which in turn increases standard errors.Yet, we still find some evidence consistent with our baseline specification. In particular, in column 3 thekey interaction term between initial political institutions and price changes still takes a negative (and 10%significant) coefficient. We have also tried a different exercise related to the timing structure, maintainingthe overlapping nature of our explanatory variable but changing the time horizon. We have thus estimatedthe effect of five and ten years rolling windows on institutional changes between t − 1and t. In the caseof the five years window, the coefficients have the same signs, but are not statistically significant; in thecase of the ten years widow, the effect is negative and significant for the average autocracy but displays noheterogeneity.

While a large majority of authors have interpreted positive values of polity2 as pertaining to democracy,one can find in the literature examples of authors who have used a more stringent criterion. Thus in Table 12

15Another plausible candidate is the variable “purges”from the Banks database, which unfortunately is not available free ofcharge.

18

we present results using alternative thresholds. Our results are statistically robust when using thresholds of 1and 2. For more demanding definitions of democracy the results are qualitatively robust, but lose statisticalsignificance. A final robustness check we performed was on the sensitivity of our results to possible outliers.We re-run our specifications excluding all the observations in the top 1% of the distribution of price changes(in absolute value) and/or in the top 1% of the distribution of polity2 changes. We also excluded all influentialobservations, as identified by the DFBETA method, once again without changes in results. These resultsare available on request.

4 Conclusions

We have presented a model of endogenous political-regime determination as a function of natural-resourcerents. The model predicts that, everything else equal, resource poor countries will be more likely to bedemocracies that resource rich ones. This is a notoriously difficult prediction to test. Hence, we use themodel to develop an additional testable implication that, we argue, better leads itself to causal identification.This prediction is that, among autocracies, resource windfalls will trigger further moves towards harsher formsof autocracy, the more so the less entrenched the autocracy was initially, while there is no impact in countriesthat start out as democracies. These predictions find empirical support in a broad panel of countries.

Future work could usefully look at other outcomes. We have briefly discussed in the text the possibilityof extending the model to deliver predictions on uses of the resource rents other than to distort the politicalrules of the game in the incumbent’s favor, such as spending on education or infrastructure. This could befurther extended to generate predictions on the growth response of the economy to resource windfalls. Itseems likely that such extensions will produce similar nonmonotonicities in the relation between resourcewindfalls and outcomes as we found in this paper, and that such predictions could be tested using a similarconditioning strategy.

The nonmonotonicites we uncover, both theoretically and empirically, imply a more nuanced policyresponse to natural-resource windfalls than has generelly been the case heretofore. While our empirical workfocuses on “local” changes in resource rents, the model predicts that a large discrete resource windfall hasthe capacity to tip a democracy into autocracy. Countries close to the democracy-autocracy threshold aretherefore more vulnerable to the impact of large resource discoveries, and should be the focus of heightenedattention from policy makers in importing countires and extractive industiers alike.

19

REFERENCES

Acemoglu, Daron, Simon Johnson, James A. Robinson, and Pierre Yared (2008). “Income and Democ-racy” American Economic Review, 98(3): 808-842.

Acemoglu, Daron, and James A. Robinson (2001). “A Theory of Political Transitions” American Eco-nomic Review, 91(4): 938-963.

Acemoglu, Daron, and James A. Robinson (2006). Economic Origins of Dictatorship and Democracy.Cambridge, UK: Cambridge University Press.

Acemoglu, Daron, James A. Robinson, and Thierry Verdier (2004). “Kleptocracy and Divide and Rule:A Model of Personal Rule” Journal of the European Economic Association, 2, 162-192.

Alexeev, Michael and Robert Conrad (2009). “The Elusive Curse of Oil” Review of Economics andStatistics, 91: 586-598.

Baland, Jean-Marie and Patrick Francois (2000). “Rent Seeking and Resource Booms” Journal of De-velopment Economics, 61 (1): 527-542.

Barro, Robert J. (1973): “The Control of Politicians: An Economic Model,” Public Choice, Volume 14,Number 1, 19-42.

Barro, Robert J. (1999). “Determinants of Democracy” Journal of Political Economy, 107(6): S158-183.Besley, Tim and Torsten Persson (2010). “The Logic of Political Violence” Quarterly Journal of Eco-

nomics, forthcoming.Besley, Tim and Masayuki Kudamatsu (2008). “Health and Democracy” American Economic Review,

Papers and Proceedings.Bond, Stephen (2002). “Dynamic Panel Data Models: a Guide to Micro Data Methods and Practice”

IFS Working Papers 09/02.Bruckner, Markus and Antonio Ciccone (2009). “Rain and the Democratic Window of Opportunity”

Econometrica, forthcoming.Bruckner, Markus and Antonio Ciccone (2010). “International Commodity Prices, Growth, and Civil

War in Sub-Saharan Africa ” Economic Journal, 120(544): 519-534.Bruckner, Markus, Antonio Ciccone, and Andrea Tesei (2011). “Oil Price Shocks, Income, and Democ-

racy” Review of Economics and Statistics, forthcoming.Burke, Paul J. and Andrew Leigh (2010). “Do Output Contractions Trigger Democratic Change?”

American Economic Journal - Macroeconomics, 2: 124-157.Cabrales, Antonio and Esther Hauk (2009). “The Quality of Political Institutions and the Curse of

Natural Resources” Economic Journal, forthcoming.Caselli, Francesco and John Coleman (2006). “On the Theory of Ethnic Conflict” unpublished, LSE.Caselli, Francesco and Tom Cunningham (2009). “Leader Behavior and the Natural Resource Curse”

Oxford Economic Papers, 61, 628-650Caselli, Francesco and Guy Michaels (2011). “Do Oil Windfalls Improve Living Standards? Evidence

from Brazil” unpublished, LSE.Cingranelli, David. L., and David L. Richards, (2008). The Cingranelli-Richards (CIRI) Human Rights

Data Project Coding Manual.Collier, Paul and Anke Hoeffler (2009). “Testing the Neocon Agenda: Democracy in Resource-Rich

Societies” European Economic Review, 53: 293-308.

20

Dalgaard, Carl-Johan and Ola Olsson (2008). “Windfall gains, Political Economy and Economic Devel-opment” Journal of African Economies, 17 (Supplement 1), pp. i72-i109.

Deaton, Angus and R. Miller (1995). International Commodity Prices, Macroeconomic Performance, andPolitics in Sub-Saharan Africa. Princeton, NJ: Princeton Studies in International Finance, No 79.

Deaton, Angus (1999). “Commodity Prices and Growth in Africa” Journal of Economic Perspectives, 13(3): 23-40.

Epstein, David L., Robert Bates, Jack Goldstone, Ida Kristensen, and Sharyn O’Halloran (2006). “Demo-cratic Transitions” American Journal of Political Science, 50(3): 551-569.

Geddes, Barbara (1999). “What Do We Know About Democratization After Twenty Years?” AnnualReview of Political Science, 2: 115-144.

Grilli, Enzo R. and Maw C. Yang (1988). “Primary Commodity Prices, Manufactured Goods Prices,and Terms of Trade of Developing Countries: What the Long Run Shows” World Bank Economic Review,2: 1-48

Haber, Stephen and Victor Menaldo (2010). “Do Natural Resources Fuel Authoritarianism? A Reap-praisal of the Resource Curse” American Political Science Review, 105 (1).

Haggard, Stephan, and Robert R. Kaufman (1997). “The Political Economy of Democratic Transitions”Comparative Politics, 29(3): 263-283.

Harvey, David I., Neil M. Kellard, Jakob B. Madsen, and Mark E. Wohar (2010). “The Prebisch-SingerHypothesis: Four Centuries of Evidence” Review of Economics and Statistics, Vol. 92(2): 367-377.

Isham, Jonathan, Michael Woolcock, Lant Pritchett, and Gwen Busby (2005). “The Varieties of ResourceExperience: Natural Resource Export Structures and the Political Economy of Economic Growth” WorldBank Economic Review, 19(2): 141-174.

Leite, Carlos and Jens Weidmann (1999). “Does Mother Nature Corrupt? Natural Resources, Corruptionand Economic Growth” IMF Working Papers 99/85.

Marshall, Monty G. and Keith Jaggers (2005). “Polity IV Project: Dataset Users’ Manual” Center forGlobal Policy, George Mason University.

Nickell, Stephen J. (1981) “Biases in Dynamic Models with Fixed Effects” Econometrica, 49(6): 1417-1426.

Olken, Benjamin, and Benjamin Jones (2009): “Hit or Miss? The Effect of Assassinations on InstitutionalChange and War,” American Economic Journal: Macroeconomics 1 (2), pp. 55-87, July.

Persson, Torsten and Guido Tabellini (2006). “Democracy and Development. The Devil in Detail”American Economic Review Papers and Proceedings, 96 (2): 319-324.

Persson, Torsten and Guido Tabellini (2009). “Democratic capital: The nexus of political and economicchange” American Economic Journal: Macroeconomics 1,88-126.

Plumper, Thomas and Neumayer, Eric (2010) “The level of democracy during interregnum periods:Recoding the polity2 score” Political analysis, 18 (2). pp. 206-226.

Roberto Rigobon, and Thomas M. Stoker (2007). “Estimation With Censored Regressors: Basic Issues”International Economic Review, vol. 48(4): 1441-1467.

Roberto Rigobon, and Thomas M. Stoker (2009). “Bias From Censored Regressors” Journal of Businessand Economic Statistics, 27(3): 340-353.

21

Ross, Michael L. (2001a). Timber Booms and Institutional Breakdown in Southeast Asia, CambridgeUniversity Press.

Ross, Michael L. (2001b). “Does Oil Hinder Democracy?” World Politics 53: 325-361.Ross, Michael L. (2009). “Oil and Democracy Revisited” Mimeo, UCLA.Sachs, Jeffrey and Andrew M. Warner (2001). “The Curse of Natural Resources” European Economic

Review, 45: 827-838.Sala-i-Martin, Xavier and Arvind Subramanian (2003). “Addressing the Natural Resource Curse: An

Illustration from Nigeria” NBER Working Paper No. 9804.Sokoloff, Kenneth L. and Stanley L. Engerman (2000). “Institutions, Factors Endowments, and Paths of

Development in the New World ” Journal of Economi Perspectives, 14(3): 217-232.Tavares, Jose’ (2003). “Does foreign aid corrupt?” Economics Letters, 79:99–106.Tornell, Aaron and Philip L. Lane (1998) “Are Windfalls a Curse? A Non-representative Agent Model

of the Current Account” Journal of International Economics, 44 (1): 83-112.Torvik, Ragnar (2002) “Natural Resources, Rent Seeking and Welfare” Journal of Development Eco-

nomics, 67: 455-470.Tsui, Kevin (2010). “More Oil, Less Democracy: Evidence from Worldwide Crude Oil Discoveries”

Economic Journal, forthcoming.Ulfelder, Jay (2007): “Natural-Resource Wealth and the Survival of Autocracy” Comparative Political

Studies, 2007 40: 995Ulfelder, Jay, and Michael Lustik (2007). “Modeling Transitions to and from Democracy” Democratiza-

tion, 14(3): 351-387.Wacziarg, Romain (2009). “The First Law of Petropolitics.” Mimeo, UCLA.Wood, Reed M. and Mark Gibney (2010). “The Political Terror Scale (PTS): A Re-introduction and a

Comparison to CIRI ” Human Rights Quarterly, 32(2): 367-400.

22

5 Appendix: Formal Analysis of the Model

5.1 Equilibria where the challenger always challenges

5.1.1 Optimal choice of B subject to challenger challenging

The problem solved by a generic incumbent when all other incumbents choose B′ and the challengeralways challenges is

maxB

β(1− Ωe−δB)V (A,B′) +A+ Θ−B,

s.t. 0 ≤ B ≤ A. Define B∗(A,B′) as the solution to

e−δB∗

=1

βΩδV (A,B′).

Since V (A,B′) > 0 B∗(A,B′) exists for every A.The equilibrium value of B must satisfy

B = 0 if B∗(A, 0) < 0

B = B∗(A,B) if 0 ≤ B∗(A,B) ≤ A

B = A if B∗(A,A) > A

Begin by computing B∗(A, 0). Since

V (A, 0) =A+ Θ

1− β(1− Ω)

we havee−δB

∗(A,0) =1− β(1− Ω)βΩδ(A+ Θ)

.

Hence B∗(A, 0) < 0 if the expression on the right hand is greater than 1. This can be rearranged to

A <1− β(1− Ω)− βΩδΘ

βΩδ≡ A0.

By PA1 in the text A0 > 0.Next we compute B∗(A,B).Since

V (A,B∗) =A+ Θ−B∗

1− β(1− Ωe−δB∗)

we have

e−δB∗

=1− β(1− Ωe−δB

∗)

βΩδ(A+ Θ−B∗).

Rearranging,

eδB∗

=βΩ [δ(A+ Θ)− 1]

1− β− βΩδ

1− βB∗.

The right side is monotonically increasing, ranging from 0 to infinity. The left side is monotonically

23

decreasing, ranging from positive to negative infinity. Hence B∗ always exists. The intercept of the left sideon the vertical axis is grater than1 if A > A0. Hence the condition B∗ > 0 is satisfied for A > A0.

Define B∗(A) the solution to the fixed point problem we have just examined. We have B∗(A0) = 0 (asthe intercept of the right side of the expression above is 1). Using the implicit function theorem we also find

B∗′(A) =1

δ(A+ Θ−B∗(A)),

B∗′′(A) = − 1δ(A+ Θ−B∗(A))2

,

so B∗(A) is increasing and concave. Plugging in the expression for A0 in B∗′(A), and recalling that B∗(A0) =0 one can check that B∗′(A0) < 1. This implies that the B∗(A) function never crosses the 45-degree line, sothe condition B∗(A,A) > A is never satisfied.

5.1.2 Optimality of decision to challenge

We have established that in this MPE (if it exists) the incumbent plays

B = 0 if A ≤ A0

B = B∗(A) if A0 < A.

Subject to the game continuing as one where challengers challenge, the challenger of a generic period chal-lenges if

βp(B)V (A,B) > Π.

Since V (A,B) is increasing in A (after taking into account the dependence of B on A, and using thefact that B∗′(A) < 1) the condition is satisfied for all values of A if it is satisfied for A = 0. For A = 0 thecondition is

βΩΘ1− β(1− Ω)

> Π,

which is PA2.

5.1.3 Absence of profitable deviations for incumbent

The last thing to check is that a generic incumbent cannot or does not wish to implement a one-perioddeviation that deters the current challenger from challenging. Call the deviation Bc(A). This is given by

Ωe−δBcβA+ Θ−B∗(A)

1− β[1− Ωe−δB∗(A)

] = Π.

Note thatΩe−δB

∗(A)βA+ Θ−B∗(A)

1− β[1− Ωe−δB∗(A)

] =1δ.

Hencee−δB

∗(A)

e−δBc(A)=

1Πδ

24

OrBc = B∗ − 1

δlog(Πδ).

The deviation dominates if

A+ Θ− Bc + βA+ Θ−B∗

1− β [1− Ωe−δB∗ ]>

A+ Θ−B∗

1− β [1− Ωe−δB∗ ]

A+ Θ−B∗ +1δ

log(Πδ) + βA+ Θ−B∗

1− β [1− Ωe−δB∗ ]>

A+ Θ−B∗

1− β [1− Ωe−δB∗ ]

A+ Θ−B∗ +(β − 1) [A+ Θ−B∗]1− β [1− Ωe−δB∗ ]

> −1δ

log(Πδ)1− β

[1− Ωe−δB

∗](A+ Θ−B∗) + (β − 1) [A+ Θ−B∗]

1− β [1− Ωe−δB∗ ]> −1

δlog(Πδ)

βΩe−δB∗

(A+ Θ−B∗)1− β [1− Ωe−δB∗ ]

> −1δ

log(Πδ)

1δ

> −1δ

log(Πδ)

1 > − log(Πδ),

which is the opposite of PA3. Hence PA3 insures the existence of a MPE where the challenger alwayschallenges. (Note that the deviation described above may be unfeasible for certain values of A in which casea fortiori the existence of the challenging equilibrium is confirmed.)

5.2 Equilibria where the challenger never challenges

Define V (A,B) the value of incumbency in a MPE where the challenger never challenges and all incum-bents play B:

V (A,B) =A+ Θ−B

1− β

The condition that assures that the challenger does not challenge is

βΩe−δBA+ Θ−B

1− β≤ Π.

For reasons already seen above there exists a unique solution to the equation

βΩe−δBA+ Θ−B

1− β= Π.

We call this solution B(A). Note that B(A) is strictly increasing.The condition that B(A) > 0 for all A is given by

βΩΘ(1− β) Π

> 1,

which is trues in view of PA2..It is obvious that any B > B cannot be an equilibrium. Any incumbent would deviate to a lower level

25

of B. Obviously for B to be an equilibrium it must be feasible, or

B(A) ≤ A.

Since B(0) > 0, there exists an interval of values for A such that an equilibrium where the challenger doesnot challenge does not exist.

Now rewrite the definition of B as

e−δB (A+ Θ−B) =Π(1− β)βΩ

,

then using the implicit function theorem

B′(A) =1

δ(A+ Θ− B(A)

)− 1

,

B′′(A) = − δ[δ(A+ Θ− B(A)

)− 1]2 < 0.

Because B(A) is concave and B(0) > 0, there exists a A > 0 such that B(A) < A for A > A. A is defined by

e−δAΘ =Π(1− β)βΩ

,

orA =

1δ

logβΩΘ

Π(1− β).

The key condition that must be satisfied by an equilibrium with no challenges is that the incumbent doesnot wish to deviate to a lower B. A generic deviation B is dominated if

A+ Θ−B + β(1− Ωe−δB)A+ Θ− B

1− β<A+ Θ− B

1− β.

The left side is maximized by B∗c (A, B), which, after rearranging the first order condition is given by

B∗c (A, B) =1δ

log δβΩV (A, B).

26

Plugging this back into the left side of the inequality above we have

A+ Θ− 1δ

log δβΩV (A, B) + β

(1− 1

δβV (A, B)

)V (A, B) <

A+ Θ− B1− β

A+ Θ− 1δ

log δβΩV (A, B) + βA+ Θ− B

1− β− 1δ

<A+ Θ− B

1− β

A+ Θ− 1δ

log δβΩV (A, B)− 1δ

< A+ Θ− B

−1δ

log δβΩV (A, B)− 1δ

< −B

1δ

[log δβΩV (A, B) + 1

]> B

exp[log δβΩV (A, B) + 1

]> eδB

Now recall that

eδB(A) =βΩV (A, B(A))

Π

so the condition for B to be an equilibrium is

exp[log(δβΩV (A, B)

)+ 1]

>βΩV (A, B)

Π

exp[log(δ) + 1 + log

(βΩV (A, B)

)]> exp

[log(1/Π) + log

(βΩV (A, B)

)]exp [log(δ) + 1] > exp [log(1/Π)]

exp [log(δ) + 1] exp − [log(1/Π)] > 1

exp [log(δ) + 1− log(1/Π)] > 1

exp [log(δ) + 1 + log(Π)] > 1

exp [log(δΠ) + 1] > 1

log(δΠ) + 1 > 0,

which once again is the opposite of PA3 so PA3 rules out equilibria where the challenger never challenges.

27

Figure 1. Theoretical relation between resource revenues and autocracy

28

Figure 2. Marginal price effect at different levels of initial Polity

All graphs plot estimates from Table 3. The upper graph plots the estimates in column (1) for the averageeffect of price changes on Polity2 changes. The central graph plots the estimates in column (2) where theeffect of price changes is conditional on the country being an autocracy or a democracy. The lower graphplots the estimates in column (3), which condition on the exact Polity2 level the country is at the start ofthe price variation.

29

Figure 3. Estimated coefficients at different bins

The graph divides the Polity scale into six bins. The bins are constructed so to maintain the symmetryaround the zero threshold, while minimizing the differences in frequency across them. For each bin theestimated coefficient of Polity2 change on price change and the relative 90% confidence intervals areplotted.

30

Table 1. Countries, by Commodity

Princ. Comm. Countries Countries

Oil 30 Algeria, Angola, Azerbaijan, Cameroon, China, Norway, Congo, Ecuador, Egypt, GabonIndonesia, Iran, Kazakhstan, Kuwait, Lybia, Malaysia, Mexico, Nigeria, Oman, Qatar,Russia, Saudi Arabia, Syria, Trinindad, Tunisia, UAE, UK, Venezuela, Vietnam, Yemen

Coffee 11 Brazil Burundi, Colombia, El Salvador, Ethiopia, Guatemala, Haiti, MadagascarNicaragua, Rwanda, Uganda

Wood 9 Austria, Canada, Estonia, Finland, Latvia, Lithuania, Portugal, Romania, Sweden

Pig Iron 8 Albania, Bhutan, Dominican Republic, Georgia, Japan, Lebanon, Slovakia, Ukraine

Gemstones 7 Armenia, Botswana, Central African Republic, India, Lesotho, Namibia, Sierra Leone

Oranges 6 Cyprus, Israel, Italy, Moldova, Spain, Turkey

Aluminum 6 Bahrain, Germany, Ghana, Mozambique, Slovenia, Switzerland

Cotton 5 Benin, Burkina Faso, Kyrgyzstan, Mali, Sudan

Bananas 4 Costa Rica, Honduras, Panama, Philippines

Beef 4 Djibouti, Ireland, New Zealand, Uruguay

Copper 4 Chile, Mongolia, Peru, Zambia

Fish 4 Bangladesh, South Korea, Malta, Tanzania

Phosphates 4 Jordan, Morocco, Senegal, Togo

Coal 3 Australia, Czech Republic, Poland

Tobacco 3 Greece, Malawi, Zimbabwe

Bauxite 2 Guinea, Jamaica

Natural gas 2 Belgium, Turkmenistan

Rice 2 Pakistan, Thailand

Swine 2 Denmark, Netherlands

Tea 2 Kenya, Sri Lanka

Wheat 2 Argentina, France

Cocoa 1 Cote d’Ivoire

Gold 1 Papua New Guinea

Groundnuts 1 Gambia

Jute 1 Nepal

Maize 1 United States

Rubber 1 Cambodia

Silver 1 South Africa

Soybean 1 Paraguay

Sugar 1 Eritrea

Tin 1 Bolivia

Uranium 1 Niger

31

Table 2. Summary Statistics.

Mean Std. Dev. Min. Max. Obs.

Avg∆Price .078 .19 -.36 1.04 5486

∆Polity .09 1.72 -18 16 5380

Polity2 .99 7.56 -10 10 5528

Polity2 1962 -.11 7.60 -10 10 4464

Polity2 2009 4.16 6.05 -10 10 6276

AvgShare .067 .09 .001 .41 6276

Countshare Princ 22.12 11.85 2 48 6276

Countshare 30.86 15.03 2 48 6276

32

Table 3. Baseline Sample

(1) (2) (3) (4)

LS LS LS SYS-GMM∆Pr -0.253

(0.17)

∆Pra -0.326* -0.418** -0.717***(0.18) (0.19) (0.22)

∆Prd 0.114 0.117 0.051(0.19) (0.18) (0.26)

∆Pra ∗ Plt−4,a -0.137** -0.139*(0.06) (0.08)

∆Prd ∗ Plt−4,d 0.041 0.030(0.06) (0.09)

Plt−4,a -0.064*** -0.073***(0.02) (0.03)

Plt−4,d -0.118*** -0.079***(0.03) (0.03)

State FE Yes Yes Yes Yes

Year FE Yes Yes Yes Yes

N. of countries 131 131 131 131

Observations 4875 4744 4744 4744

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is

the t−1 to t change in the revised Polity score (Polity2). The method of estimation in columns

(1)-(3) is least squares, in column (4) system-GMM (Blundell-Bond). The values in brackets

are Huber robust standard errors clustered at the country level.

33

Table 4. Export Shares

(1) (2) (3) (4) (5) (6)

Above 1st Decile Share Above 1st Quartile Share Above Median Share

∆Pra -0.268 -0.374* -0.278 -0.407* -0.407 -0.494*(0.20) (0.20) (0.21) (0.21) (0.26) (0.26)

∆Prd 0.289 0.271 0.350 0.328 0.435 0.402(0.19) (0.18) (0.23) (0.22) (0.31) (0.31)

∆Pra ∗ Plt−4,a -0.135** -0.162*** -0.112**(0.06) (0.05) (0.05)

∆Prd ∗ Plt−4,d 0.023 0.058 -0.048(0.07) (0.07) (0.09)

Plt−4,a -0.078*** -0.074*** -0.060**(0.03) (0.02) (0.03)

Plt−4,d -0.125*** -0.138*** -0.071*(0.03) (0.03) (0.04)

State FE Yes Yes Yes Yes Yes Yes

Year FE Yes Yes Yes Yes Yes Yes

N. of countries 117 117 93 93 63 63

Observations 4206 4206 3413 3413 2312 2312

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t− 1

to t change in the revised Polity score (Polity2). The values in brackets are Huber robust standard

errors clustered at the country level. Columns (1)-(2) exclude countries in the first decile of average

share for the principal commodity. Columns (3)-(4) exclude countries in the first quartile of average

share for the principal commodity. Columns (5)-(6) exclude countries below the median of average

share for the principal commodity.

34

Table 5. Shares Availability

(1) (2) (3) (4) (5) (6)

Above 1st Quartile Obs. Observations before 1986 Always 1st after 1986

∆Pra -0.360* -0.500** -0.393** -0.515** -0.321* -0.428**(0.21) (0.21) (0.20) (0.20) (0.18) (0.19)

∆Prd 0.214 0.203 0.134 0.132 0.116 0.124(0.20) (0.19) (0.21) (0.20) (0.20) (0.19)

∆Pra ∗ Plt−4,a -0.163** -0.158** -0.142**(0.06) (0.06) (0.06)

∆Prd ∗ Plt−4,d 0.007 0.037 0.036(0.08) (0.08) (0.08)

Plt−4,a -0.070** -0.071*** -0.070***(0.03) (0.03) (0.03)

Plt−4,d -0.105*** -0.116*** -0.113***(0.03) (0.03) (0.03)

State FE Yes Yes Yes Yes Yes Yes

Year FE Yes Yes Yes Yes Yes Yes

N. of countries 96 96 92 92 105 105

Observations 3795 3795 3724 3724 4072 4072

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t−1 to

t change in the revised Polity score (Polity2). The values in brackets are Huber robust standard errors

clustered at the country level. Columns (1)-(2) exclude countries in the first quartile of observations

on which the principal commodity is identified. Columns (3)-(4) exclude countries that don’t have

observations for share before 1986, the midpoint year of share observations. Columns (5)-(6) include

countries that don’t have observations for share before 1986, but whose principal commodity has always

been ranked first afterwards.

35

Table 6. Boundary Observations

(1) (2) (3) (4)

∆Pl ≥ 0 D=1|∆Pl ≥ 0 Unbounded Countries Unbounded Obs.

∆Pra -0.448** -0.052 -0.391* -0.413*(0.17) (0.03) (0.23) (0.21)

∆Prd -0.021 -0.009 0.075 0.106(0.12) (0.03) (0.29) (0.26)

∆Pra ∗ Plt−4,a -0.158*** -0.028** -0.134* -0.139*(0.05) (0.01) (0.08) (0.07)

∆Prd ∗ Plt−4,d 0.032 0.012 0.076 0.032(0.04) (0.01) 0.076 (0.11)

Plt−4,a -0.013 -0.004 -0.080*** -0.076***(0.02) (0.00) (0.03) (0.03)

Plt−4,d -0.075*** -0.013*** -0.150*** -0.133***(0.02) (0.00) (0.03) (0.03)

State FE Yes Yes Yes Yes

Year FE Yes Yes Yes Yes

N. of countries 131 131 90 107

Observations 4592 4744 3222 3634

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t− 1 to t change in the

revised Polity score (Polity2). The values in brackets are Huber robust standard errors clustered at the country level.

Column (1) considers non-negative Polity2 changes only. Column (2) estimates a dummy variable which takes value of

1 if there is a positive Polity2 change, and 0 otherwise. Column (3) restricts the sample to countries that never touched

the boundaries at -10 and +10 on the Polity scale. Column (4) excludes the observations at -10 and +10.

36

Table 7. Big Producers

(1) (2) (3) (4)

Exclude OPEC countries Exclude Big Producers

∆Pra -0.361 -0.433** -0.428* -0.536**(0.22) (0.21) (0.26) (0.24)

∆Prd 0.087 0.092 -0.016 -0.049(0.20) (0.19) (0.26) (0.25)

∆Pra ∗ Plt−4,a -0.147** -0.153*(0.07) (0.08)

∆Prd ∗ Plt−4,d 0.036 0.074(0.06) (0.08)

Plt−4,a -0.066** -0.084***(0.03) (0.03)

Plt−4,d -0.120*** -0.130***(0.03) (0.04)

State FE Yes Yes Yes Yes

Year FE Yes Yes Yes Yes

N. of countries 120 120 87 87

Observations 4282 4282 3003 3003

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable

is the t − 1 to t change in the revised Polity score (Polity2). The values in brackets are

Huber robust standard errors clustered at the country level. Columns (1)-(2) exclude OPEC

countries. Columns (3)-(4) exclude countries producing more than 3% of total world pro-

duction in their principal commodity. Detail on the sources used to identify big producers

are reported in Table 7b.

37

Table 8. Alternative Definitions Price Index.

(1) (2) (3) (4) (5)

Average 1975 1980 1990 2001

∆Indexa -0.461 -0.491* -0.461 -0.515* -0.518*(0.33) (0.28) (0.28) (0.30) (0.28)

∆Indexd 0.176 0.318 0.299 0.104 0.304(0.35) (0.38) (0.22) (0.37) (0.33)

∆Indexa ∗ Plt−4,a -0.097 -0.118 -0.157* -0.220** -0.111(0.09) (0.08) (0.09) (0.12) (0.08)

∆Indexd ∗ Plt−4,d 0.071 0.117 -0.007 0.011 0.083(0.09) (0.10) (0.09) (0.14) (0.08)

Plt−4,a -0.067*** -0.074** -0.064** -0.025 -0.072***(0.02) (0.03) (0.03) (0.03) (0.02)

Plt−4,d -0.098*** -0.099*** -0.064 -0.055 -0.096***(0.03) (0.04) (0.04)) (0.04) (0.03)

State FE Yes Yes Yes Yes Yes

Year FE Yes Yes Yes Yes Yes

Observations 4957 3600 3527 3327 4629

Countries 131 84 82 78 122

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t− 1 to t change

in the revised Polity score (Polity2). The values in brackets are Huber robust standard errors clustered at the

country level. All columns follow the methodology in Deaton and Miller (1995) to construct the weighted price

index of commodities. We use all commodities included in the UN COMTRADE database and whose price series

is identified in the IMF IFS database. Columns (1) weights commodities by their average share in country exports,

measured over all available years. Column (2) weights commodities by their share in 1975, the base year used in

Deaton-Miller (1995). Column (3) weights commodities by their share in 1980, the base year used in Besley-Persson

(2008). Column (4) weights commodities by their share in 1990, the base year used in Deaton (1999) and Bruckner

and Ciccone (2010). Column (5) weights commodities by their share in 2001, the year with the highest number of

reporting countries.

38

Table 9. Typologies of Commodities

(1) (2) (3) (4) (5) (6) (7) (8)

Point Source Diffuse Mineral Non Mineral

∆Pra -0.273 -0.372* -1.583* -1.633* -0.141 -0.222 -1.062** -1.069***(0.21) (0.21) (0.82) (0.82) (0.29) (0.29) (0.41) (0.40)

∆Prd 0.250 0.237 -0.511 -0.436 0.236 0.224 -0.225 -0.239(0.22) (0.22) (0.46) (0.44) (0.28) (0.27) (0.36) (0.33)

∆Pra ∗ Plt−4,a -0.177*** 0.331 -0.119** -0.037(0.05) (0.57) (0.05) (0.21)

∆Prd ∗ Plt−4,d 0.021 0.138 0.064 -0.008(0.07) (0.13) (0.07) (0.13)

Plt−4,a -0.051** -0.116 -0.079** -0.053(0.02) (0.08) (0.02) (0.05)

Plt−4,d -0.111*** -0.126*** -0.123*** -0.109***(0.03) (0.05) (0.03) (0.04)

State FE Yes Yes Yes Yes Yes Yes Yes Yes

Year FE Yes Yes Yes Yes Yes Yes Yes Yes

N. of countries 87 87 44 44 70 70 61 61

Observations 3127 3127 1617 1617 2494 2494 2250 2250

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t− 1 to t change in the

revised Polity score (Polity2). The values in brackets are Huber robust standard errors clustered at the country level.

Mineral commodities are: Aluminum, Bauxite, Coal, Copper, Gemstones, Gold, Natural Gas, Oil, Pig Iron, Phosphates,

Silver, Tin, Uranium. Non-mineral commodities are: Bananas, Beef Meat, Cocoa, Coffee, Cotton, Fish, Groundnuts,

Jute, Maize, Oranges, Rice, Rubber, Soybean, Sugar, Swine Meat, Tea, Tobacco, Wheat, Wood. The point source/diffuse

distinction follows almost the same classification, but assign countries producing Bananas, Cocoa, Coffee and Sugar to

the other category.

39

Table 10. Exclude Transition Periods

(1) (2) (3)

∆Pr -0.074(0.15)

∆Pra -0.085 -0.171(0.18) (0.19)

∆Prd 0.192 0.195(0.18) (0.17)

∆Pra ∗ Plt−4,a -0.123**(0.06)

∆Prd ∗ Plt−4,d 0.020(0.07)

Plt−4,a -0.035(0.02)

Plt−4,d -0.081***(0.03)

State FE Yes Yes Yes

Year FE Yes Yes Yes

N. of countries 131 131 131

Observations 4751 4555 4555

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent

variable is the t − 1 to t change in the revised Polity score (Polity2). The values in

brackets are Huber robust standard errors clustered at the country level. Columns

(1) to (3) consider the baseline sample of 131 countries and replace transition years

(Polity=-77) by missing values.

40

Table 11. Alternative Timing Structure: 3 years non-overlap

(1) (2) (3)

∆Pr 0.082(0.14)

∆Pra 0.023 -0.042(0.13) (0.15)

∆Prd 0.239 0.228(0.15) (0.14)

∆Pra ∗ Plt−4,a -0.078*(0.05)

∆Prd ∗ Plt−4,d -0.026(0.05)

Plt−4,a -0.276***(0.07)

Plt−4,d -0.310***(0.09)

State FE Yes Yes Yes

Year FE Yes Yes Yes

Observations 1621 1599 1599

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent

variable is the t − 3 to t change in the revised Polity score (Polity2). The values

in brackets are Huber robust standard errors clustered at the country level. The

estimation is for three-years non-overlapping periods: the change in Polity2 in period

1963-1966 is explained by the change in price in period 1962-1965; the change in

Polity2 in period 1966-1969 is explained by the change in price in period 1965-1968;

and similarly for following periods.

41

Table 12. Alternative Thresholds for Democracy

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)

Polity ≥ 1 Polity ≥ 2 Polity ≥ 3 Polity ≥ 4 Polity ≥ 5

∆Pra -0.353* -0.484*** -0.360* -0.458** -0.362* -0.408** -0.308* -0.296* -0.363* -0.289(0.18) (0.18) (0.18) (0.18) (0.18) (0.18) (0.18) (0.18) (0.19) (0.18)

∆Prd 0.198 0.175 0.176 0.160 0.147 0.149 0.094 0.112 0.071 0.131(0.19) (0.18) (0.19) (0.18) (0.19) (0.18) (0.20) (0.19) (0.19) (0.19)

∆Pra ∗ Plt−4,a -0.164*** -0.125** -0.080 -0.012 -0.004(0.06) (0.05) (0.05) (0.04) (0.03)

∆Prd ∗ Plt−4,d -0.038 -0.024 -0.007 0.057 0.065(0.06) (0.07) (0.06) (0.11) (0.15)

Plt−4,a -0.62*** -0.064*** -0.069*** -0.071*** -0.082***(0.03) (0.01) (0.02) (0.01) (0.01)

Plt−4,d -0.109*** -0.104*** -0.105*** -0.087** -0.114***(0.03) (0.04) (0.03) (0.04) (0.04)

State FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Year FE Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes

Observations 4744 4744 4744 4744 4744 4744 4744 4744 4744 4744

∗∗∗,∗∗ ,∗ represent significance at 1%, 5%, and 10%, respectively. The dependent variable is the t − 1 to t change in

the revised Polity score (Polity2). The values in brackets are Huber robust standard errors clustered at the country

level.

42

Appendix Table A1

(1) (2) (3) (4) (5) (6)

Plt−4 #∆Pl ≤ 0 #(Pl = −10|∆Pl ≤ 0) % to Pl=-10 #∆Pl ≥ 0 #(Pl = 10|∆Pl ≥ 0) % to Pl=10

-10 4 1 0.25 9 0 0

-9 5 2 0.4 40 0 0

-8 5 2 0.4 18 0 0

-7 16 0 0 76 0 0

-6 8 0 0 29 0 0

-5 6 0 0 17 0 0

-4 5 0 0 10 0 0

-3 7 0 0 7 0 0

-2 2 0 0 9 0 0

-1 7 0 0 13 0 0

0 4 0 0 3 0 0

1 4 0 0 5 0 0

2 3 0 0 6 0 0

3 4 0 0 6 0 0

4 5 0 0 5 0 0

5 10 0 0 13 0 0

6 12 0 0 20 0 0

7 16 0 0 17 1 0.059

8 19 0 0 17 1 0.059

9 13 0 0 15 9 0.6

10 5 0 0 1 0 0

Column (1) reports the number of negative changes in the Polity2 score at each initial level of Polity2 at t-4. Column (2) reports the

number of negative changes that bring the Polity2 score at -10. Column (3) calculates the percentage. Column (4) reports the number

of positive changes in the Polity2 score at each initial level of Polity2 at t-4. Column (5) reports the number of positive changes that

bring the Polity2 score at 10. Column (6) calculates the percentage.

43

Appendix Table A2. Big Producers, by Commodity

Commodity Countries

Oil Algeria, Angola, China, Indonesia, Iran, Kuwait, Lybia, MexicoNorway, Russia, Saudi Arabia, UAE, Venezuela

Coffee Brazil, Colombia, Ethiopia, Guatemala

Wood Canada, Finland, Sweden

Tea Kenya, Sri Lanka

Bananas Costa Rica, Philippines

Oranges Italy, Spain

Copper Chile, Peru, Zambia

Bauxite Guinea, Jamaica

Phosphates Jordan, Morocco

Uranium Niger

Tobacco Malawi

Rice Thailand

Cotton Mali

Coal Australia

Cocoa Cote d’Ivoire

Maize United States

Beef France

Gemstones Botswana

Pig Iron Ukraine

Tin Bolivia

Data for commodities produced in a country that constitute more than 3% of total world supply

are obtained from the following sources: Copper, Gold, Bauxite, Tin, Phosphates, Uranium, Gem-

stones (British Geological Survey 2000-2008, available here); Cocoa, Bananas, Oranges, Beef, Jute,

Maize, Wood, Rice, Sugar, Tea, Tobacco (Food and Agricultural Organization 1970-2009, available

here); Coffee (International Coffee Organization 1980-2009, available here); Cotton (US Department of

Agriculture 1970-2009, available here); Coal, Oil (US Energy Information Administration 1980-2009,

available here).

44